SPE-YOLO: A deep learning model focusing on small pulmonary embolism detection

Area of Science:

• Artificial Intelligence in Medical Imaging
• Deep Learning for Diagnostic Support
• Pulmonary Embolism Detection

Background:

• Small pulmonary embolisms pose diagnostic challenges due to their subtle presentation.
• Accurate and timely detection of pulmonary embolism is crucial for effective patient management.
• Existing deep learning models require optimization for detecting small-sized pathologies.

Purpose of the Study:

• To develop and validate the SPE-YOLO deep learning model for enhanced detection of small pulmonary embolisms.
• To improve the accuracy and efficiency of pulmonary embolism diagnosis using artificial intelligence.
• To provide clinicians with a more reliable tool for identifying subtle pulmonary embolism cases.

Main Methods:

• A retrospective study utilized the YOLOv8 framework, incorporating a P2 detection head for small targets.
• The SEAttention mechanism was integrated to focus on critical features, improving detection accuracy.
• ODConv convolution was introduced for refined feature extraction, capturing comprehensive contextual information.

Main Results:

• SPE-YOLO achieved 84.20% precision and 81.50% accuracy on the Tianjin test set, outperforming original YOLOv8.
• Improvements were observed in F1 scores (4%), recall rates (5%), and average accuracy (6%).
• External validation on the RSNA dataset showed 90.70% sensitivity and 86.45% accuracy, demonstrating robust performance.

Conclusions:

• The SPE-YOLO algorithm effectively identifies small pulmonary embolisms, offering enhanced diagnostic capabilities.
• This AI-driven approach provides clinicians with a more accurate and efficient tool for pulmonary embolism diagnosis.
• The developed model is poised to advance the quality of pulmonary embolism diagnostics and medical services.